Image recognition technology using deep learning is being actively researched every day, and the application fields of image recognition technology are expanding. For example, in the medical field, which many images are collected and diagnosed by humans, and in the agricultural field, which relies heavily on human experience. In this paper, the detection of lesion candidates in capsule endoscope images is described as an example of application of image recognition technology using deep learning to the medical field. And, as an example of application in the agricultural field, the anomaly detection in our automatic cherry sorting system is presented.
In this study, to realize a simple identification of surgical steel instruments, we proposed a method to detect surgical steel instruments using Faster R-CNN on ortho-images of surgical steel instruments acquired with an RGB-D camera, and furthermore, to identify surgical steel instruments based on their dimensional information. In the proposed method, ortho-images are generated by orthographic transformation of 3D point clouds acquired using Intel REALSENSE DEPTH CAMERA D435i. A faster R-CNN object detector was applied to the ortho-image to detect surgical steel instruments. Since the ortho-image of the proposed method contains dimensional information, the area value of the region containing the surgical steel instrument can be easily calculated, and the surgical steel instrument can be identified without image scale correction.
The surface SH-wave acoustoelasticity is one of the acoustoelastic methods, which is a stress measurement method based on the phenomenon that the propagation velocity of elastic waves changes depending on the stress state. Surface SH-wave acoustoelasticity can measure the principal stress difference from the difference in propagation velocities of surface SH waves propagating in two orthogonal directions, but it cannot measure the principal stress itself. However, if the propagation velocity of the surface SH-wave in the stress-free state is known, it is possible to measure the principal stress in the surface SH-wave acoustoelastic method. Therefore, in this study, we attempted to measure the principal stress using the surface SH-wave acoustoelasticity by estimating the propagation velocity in the stress-free state using surface anisotropy, which is a stress-independent parameter of the surface structure. Experimental results demonstrated the possibility of using surface anisotropy to estimate the propagation velocity in the stress-free state and measure the principal stress within ±10 MPa as well as the principal stress difference.
We propose a system targeting visually impaired individuals that provides a couple of functions that help what they claim they are always in trouble when going out using public transportations. One of the functions alerts users when the bus that they are planning to ride gets closer to the bus stop where they are waiting for the bus. It runs as a multi-thread system. One of the threads periodically obtains the location of the bus driver utilizing a GPS receiver and uploads it to a specific shared location on the web. And the other thread fetches the stored shared location from the web and calculates the distance to the location of the bus stop where the user is waiting for the bus. The latter thread alerts the user when the distance gets lower than a threshold. The other function we propose alerts users when their intended bus stop is approaching within a certain distance during bus rides. The system utilizes the GPS functionality embedded in smartphones to acquire the user's latitude and longitude. By comparing this information with the coordinates pre-registered in the system for each bus stop, it detects when the user reaches within a certain distance to the destination bus STOP. The implementation of the proposed system demonstrated the intended functionality in an experiment.
PaDiM, an anomaly detection model using a pre-trained CNN on ImageNet, shows high performance. However, the pre-trained CNN model has a texture bias, resulting in poor performance for global anomalies. Therefore, we used Fast Fourier Convolution (FFC) to extract global features by extracting features from Fourier space in addition to feature extraction using 3 × 3 convolutional filter in the pre-trained model. Total Variation regularization was applied to the feature map of the FFC block during pre-training. This improved anomaly detection performance for global anomalies and robustness to perturbations in the frequency band.
This research introduces real-time monitoring and localizing product stock using the First-In-First-Out (FIFO) method with radio frequency identification (RFID) pressure sensing tags. The proposed FIFO system has RFID pressure sensing tags that detect pressure changes, generate sensor codes, and estimate a stock location. Integrating artificial intelligence (AI) improves location estimation accuracy of up-to-date stock visibility. The ability of RFID pressure sensing tags and AI to detect stock from sensor code data enables more effective asset management, optimizes stock management, reduces product waste, and contributes to better efficiency in product inventory management.
In this study, visual assist signals of unilateral spatial neglect (USN) were created from Kinsbourne’s hypothesis and implemented in an inspection system. While conventional inspection is a fixed view of the patient’s ability, the developed system enables measurement of a variation in the ability of USN patients with and without visual assist signals. Experimental results using the developed system showed that visual assist signals could alleviate USN and other attention-disorder behaviors but that it is necessary to select visual assist signals according to the patient’s symptoms. In addition, the difference between the presence and absence of the visual assist signal made it possible to measure and visualize zone of proximal development, which had been treated conceptually. In the future, it is expected that the developed visual assist signals implemented in augmented reality (AR) devices will improve the daily activities of USN patients.
In recent years, the focus on mental health care in Japan has increased, leading to more attention from companies towards the mental well-being of their employees. Chronic stress, stemming from various sources including work and relationships, poses a significant threat to physical health. To address this, there is a growing need for daily monitoring of stress levels and early detection of chronic stress. This paper presents a preliminary experimental study into the relationship between facial skin temperature distribution and chronic stress. The study investigates the correlation between subjective measures obtained through questionnaires and facial skin temperature distribution, hypothesizing that chronic stress can be estimated based on FTI. The study utilizes both psychological questionnaires and FTI to examine the relationship between subjective and physiological indicators of stress. Results indicate correlations between certain components of FTI, such as asymmetry and facial regions, and psychological stress levels as measured by the questionnaire. However, further analysis reveals complexities in these relationships, suggesting that certain features in FTI may not directly correspond to psychological stress. Limitations of the study, including misalignment of feature points and variations in stress events among subjects, are discussed.
In the dairy farming sector, the need for efficient individual cow management via ICT has intensified owing to a declining workforce in the industry and a simultaneous increase in livestock numbers. To address challenges such as reducing night-time cow-monitoring hours for workers and preventing calving accidents, we developed a system capable of automatically detecting straining during labor (a key indicator of impending cow calving). Our approach involved collecting waveform data on cow movements from an unrestrained cow positioned on a sensor sheet. The collected data were subsequently analyzed using deep learning techniques. Employing a sample of 40 cows, veterinarians correlated data collected using the sensor sheet with those collected using a video camera, classifying the data into straining and other movements unrelated to calving. This curated dataset was then used to train a staining detection model using a convolutional neural network, the accuracy of which was verified. Consequently, we successfully used the staining detection model to predict cow calving with a high accuracy rate of > 95%.
In closed spaces after disasters, information gathering using mobile robots is conducted to prevent secondary disasters, such as injuries or fatalities of rescue teams. However, in such closed spaces, it is necessary for the mobile robots to establish their own communication infrastructure. Therefore, a method has been considered where multi-stage relay devices capable of relaying wireless LAN are deployed to expand the range of tele-operation. Nevertheless, there is a risk that the wireless LAN may be disconnected due to multi-stage relay device malfunctions during the mission, causing the mobile robot to become an obstacle and hindering information gathering. To address this issue, we have proposed a method when, in the event of communication disconnection with the mobile robot operating through wireless LAN, LoRaWAN is used as a sub communication infrastructure to enable retreat. However, LoRaWAN is characterized by one-way communication from the child device to the parent device. Consequently, when performing tele-operation through multi-stage relay using LoRaWAN, there is a problem of accumulated delays during continuous packet transmission. In this paper, we propose a low-latency multi-stage relay tele-operation method using LoRaWAN and demonstrate its effectiveness.
In virtual reality, it is necessary to clarify what kind of auditory stimuli are effective in presenting a pseudo-haptics sensation through auditory stimulation by cross-modal phenomena. In this study, we varied the pitch of the metallic and wooden sounds produced when a dumbbell impacted the floor. Subjects’ impressions of the impact sound and the weight of the dumbbells were recorded during use of the dumbbells using sensitivity and weight ratings. For both tones, the impressions received from the images containing the sound stimuli were often “quietness”, “intensity”, “vagueness” and “dullness”, suggesting that the reference sound was perceived as heavier when the pitch of the sound containing frequencies below 1 kHz was reduced.
The web service “Tame-Map” enables users to easily transmit, view, and share information about small local activities, known as micro-events, which typically receive less attention on mainstream social media platforms. This study aims to develop a method for predicting the click rates of micro-event announcement images on “Tame-Map” and to evaluate the effectiveness of this approach. The research primarily utilizes machine learning techniques and constructs a predictive model employing Convolutional Neural Networks (CNN). The model assumes that thumbnail images of micro-events possess distinct features that attract viewers and utilizes these images as primary input data. Moreover, the study introduces an advanced model that incorporates meta-information, such as the event's location, as additional explanatory variables. The resulting model demonstrates adequate predictive accuracy, suggesting its potential as a standard method in future applications.
This paper presents a CSI-based human detection system with commodity sensors: M5Stack. Especially, we remedy too sensitivity to individual with respect to measured CSI. Monitoring systems considering privacy proof is important for monitoring the elderly and children. In this paper, we carry out experiments on indoor human location estimation using Channel State Information (CSI). Specifically, to improve recognition accuracy in cases where training data and test data are generated from different persons, we proposed a method that includes dimensionality reduction using Principal Component Analysis (PAC) as a preprocessing. In our experiments, we clarify problems of CSI-based method and solve the problems using the proposed method. In the evaluation experiments, we achieved an improvement of 10% to 30% depending on individuals.